GR00T-N1.7-monster-tray-pickplace

GR00T N1.7 — Monster Tray Pick-and-Place (Unitree G1)

A NVIDIA GR00T N1.7 (3B) checkpoint fine-tuned for a dual-arm humanoid pick-and-place task on the Unitree G1: pick up a monster energy drink can and place it on a tray. Covers two object variants, each with its own language instruction:

• pick up the green monster can and place it on the tray.
• pick up the orange monster can and place it on the tray.

Training

• Base model: NVIDIA GR00T N1.7 (3B parameters)
• Training data: tysyuvraj/monster-tray-pickplace — teleoperated demonstrations on the same G1 hardware, dual-arm + Dex1 grippers, 2 cameras (head + wrist)
• Training length: 30,000 steps
• Final training loss: ~0.008
• Embodiment tag: NEW_EMBODIMENT (Isaac-GR00T convention for a custom, non-reference embodiment)

Intended Use

Serving as the manipulation policy ("body") in a real-time control loop, optionally supervised by an external monitor that can start/stop/redirect the policy based on human intent (this checkpoint was developed as part of a research system doing exactly that — an anticipatory, multimodal-LLM-based HRI monitor supervising this policy). It is not intended as a general can-grasping or general humanoid-manipulation model — training data covers exactly two objects and one receptacle.

How to Serve

Using Isaac-GR00T's reference eval server:

python gr00t/eval/run_gr00t_server.py \
    --embodiment_tag NEW_EMBODIMENT \
    --model_path tysyuvraj/GR00T-N1.7-monster-tray-pickplace \
    --host 0.0.0.0 --port 5555

Known Limitations

• Object-dependent placement reliability. The orange-can task completes reliably. The green-can task is less reliable at the final placement step — observed failure modes include the policy stalling in an approach/grasp loop without placing, and in some cases placing the can and then re-grasping it. This is a motor-competence / training-data gap, not a perception or control-loop issue on the deployment side. Additional green-can placement demonstrations are the most direct fix.
• Gripper calibration dependency. This checkpoint was trained against a specific dex1 gripper calibration (open/close DDS command range). If the gripper hardware or its calibration changes, the policy's raw gripper output must be remapped into the new range before being sent to the motor — sending it unmodified will silently produce near-zero effective closure even though the policy's underlying grasping behavior is correct.
• Evaluated only on the exact G1 hardware configuration used for data collection (same camera mounts, same gripper hardware revision).

Citation

@misc{singh2026grootn17monstertray,
  author = {Singh, Yuvraj},
  title  = {GR00T N1.7 -- Monster Tray Pick-and-Place (Unitree G1)},
  year   = {2026},
  url    = {https://huggingface.co/tysyuvraj/GR00T-N1.7-monster-tray-pickplace}
}